Upload 16 files

README.md

@@ -1,3 +1,126 @@
 ---
+language:
+- uz
 license: apache-2.0
+tags:
+- automatic-speech-recognition
+- generated_from_trainer
+- hf-asr-leaderboard
+- mozilla-foundation/common_voice_8_0
+- robust-speech-event
+datasets:
+- mozilla-foundation/common_voice_8_0
+base_model: facebook/wav2vec2-xls-r-300m
+model-index:
+- name: XLS-R-300M Uzbek CV8
+  results:
+  - task:
+      type: automatic-speech-recognition
+      name: Automatic Speech Recognition
+    dataset:
+      name: Common Voice 8
+      type: mozilla-foundation/common_voice_8_0
+      args: uz
+    metrics:
+    - type: wer
+      value: 15.065
+      name: Test WER (with LM)
+    - type: cer
+      value: 3.077
+      name: Test CER (with LM)
+    - type: wer
+      value: 32.88
+      name: Test WER (no LM)
+    - type: cer
+      value: 6.53
+      name: Test CER (no LM)
 ---
+
+# XLS-R-300M Uzbek CV8
+
+This model is a fine-tuned version of [facebook/wav2vec2-xls-r-300m](https://huggingface.co/facebook/wav2vec2-xls-r-300m) on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - UZ dataset.
+It achieves the following results on the validation set:
+- Loss: 0.3063
+- Wer: 0.3852
+- Cer: 0.0777
+
+## Model description
+
+For a description of the model architecture, see [facebook/wav2vec2-xls-r-300m](https://huggingface.co/facebook/wav2vec2-xls-r-300m)
+
+The model vocabulary consists of the [Modern Latin alphabet for Uzbek](https://en.wikipedia.org/wiki/Uzbek_alphabet), with punctuation removed. 
+Note that the characters <‘> and <’> do not count as punctuation, as <‘> modifies \<o\> and \<g\>, and <’> indicates the glottal stop or a long vowel.
+
+The decoder uses a kenlm language model built on common_voice text.
+
+## Intended uses & limitations
+
+This model is expected to be of some utility for low-fidelity use cases such as:
+- Draft video captions
+- Indexing of recorded broadcasts
+
+The model is not reliable enough to use as a substitute for live captions for accessibility purposes, and it should not be used in a manner that would infringe the privacy of any of the contributors to the Common Voice dataset nor any other speakers.
+
+## Training and evaluation data
+
+The 50% of the `train` common voice official split was used as training data. The 50% of the official `dev` split was used as validation data, and the full `test` set was used for final evaluation of the model without LM, while the model with LM was evaluated only on 500 examples from the `test` set.
+
+The kenlm language model was compiled from the target sentences of the train + other dataset splits.
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 3e-05
+- train_batch_size: 32
+- eval_batch_size: 8
+- seed: 42
+- gradient_accumulation_steps: 4
+- total_train_batch_size: 128
+- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
+- lr_scheduler_type: linear
+- lr_scheduler_warmup_steps: 500
+- num_epochs: 100.0
+- mixed_precision_training: Native AMP
+
+### Training results
+
+| Training Loss | Epoch | Step  | Validation Loss | Wer    | Cer    |
+|:-------------:|:-----:|:-----:|:---------------:|:------:|:------:|
+| 3.1401        | 3.25  | 500   | 3.1146          | 1.0    | 1.0    |
+| 2.7484        | 6.49  | 1000  | 2.2842          | 1.0065 | 0.7069 |
+| 1.0899        | 9.74  | 1500  | 0.5414          | 0.6125 | 0.1351 |
+| 0.9465        | 12.99 | 2000  | 0.4566          | 0.5635 | 0.1223 |
+| 0.8771        | 16.23 | 2500  | 0.4212          | 0.5366 | 0.1161 |
+| 0.8346        | 19.48 | 3000  | 0.3994          | 0.5144 | 0.1102 |
+| 0.8127        | 22.73 | 3500  | 0.3819          | 0.4944 | 0.1051 |
+| 0.7833        | 25.97 | 4000  | 0.3705          | 0.4798 | 0.1011 |
+| 0.7603        | 29.22 | 4500  | 0.3661          | 0.4704 | 0.0992 |
+| 0.7424        | 32.47 | 5000  | 0.3529          | 0.4577 | 0.0957 |
+| 0.7251        | 35.71 | 5500  | 0.3410          | 0.4473 | 0.0928 |
+| 0.7106        | 38.96 | 6000  | 0.3401          | 0.4428 | 0.0919 |
+| 0.7027        | 42.21 | 6500  | 0.3355          | 0.4353 | 0.0905 |
+| 0.6927        | 45.45 | 7000  | 0.3308          | 0.4296 | 0.0885 |
+| 0.6828        | 48.7  | 7500  | 0.3246          | 0.4204 | 0.0863 |
+| 0.6706        | 51.95 | 8000  | 0.3250          | 0.4233 | 0.0868 |
+| 0.6629        | 55.19 | 8500  | 0.3264          | 0.4159 | 0.0849 |
+| 0.6556        | 58.44 | 9000  | 0.3213          | 0.4100 | 0.0835 |
+| 0.6484        | 61.69 | 9500  | 0.3182          | 0.4124 | 0.0837 |
+| 0.6407        | 64.93 | 10000 | 0.3171          | 0.4050 | 0.0825 |
+| 0.6375        | 68.18 | 10500 | 0.3150          | 0.4039 | 0.0822 |
+| 0.6363        | 71.43 | 11000 | 0.3129          | 0.3991 | 0.0810 |
+| 0.6307        | 74.67 | 11500 | 0.3114          | 0.3986 | 0.0807 |
+| 0.6232        | 77.92 | 12000 | 0.3103          | 0.3895 | 0.0790 |
+| 0.6216        | 81.17 | 12500 | 0.3086          | 0.3891 | 0.0790 |
+| 0.6174        | 84.41 | 13000 | 0.3082          | 0.3881 | 0.0785 |
+| 0.6196        | 87.66 | 13500 | 0.3059          | 0.3875 | 0.0782 |
+| 0.6174        | 90.91 | 14000 | 0.3084          | 0.3862 | 0.0780 |
+| 0.6169        | 94.16 | 14500 | 0.3070          | 0.3860 | 0.0779 |
+| 0.6166        | 97.4  | 15000 | 0.3066          | 0.3855 | 0.0778 |
+
+
+### Framework versions
+
+- Transformers 4.16.2
+- Pytorch 1.10.2+cu102
+- Datasets 1.18.3
+- Tokenizers 0.11.0

mozilla-foundation_common_voice_8_0_uz_test[_500]_eval_results.txt

@@ -0,0 +1,2 @@
+WER: 0.150650789255054
+CER: 0.03076592082616179

mozilla-foundation_common_voice_8_0_uz_test_eval_results.txt

@@ -0,0 +1,2 @@
+WER: 0.3288115957647439
+CER: 0.06534626547372732

preprocessor_config.json

@@ -0,0 +1,10 @@
+{
+  "do_normalize": true,
+  "feature_extractor_type": "Wav2Vec2FeatureExtractor",
+  "feature_size": 1,
+  "padding_side": "right",
+  "padding_value": 0,
+  "processor_class": "Wav2Vec2ProcessorWithLM",
+  "return_attention_mask": true,
+  "sampling_rate": 16000
+}

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa3ebb6bf2207974a80f51af9ea129771bc599abed9a0f00f4b93e0bf058b624
+size 1262058993

requirements.txt

@@ -0,0 +1,10 @@
+unidecode
+tensorboard
+torch
+torchaudio
+jiwer
+soundfile
+transformers
+datasets
+pyctcdecode
+https://github.com/kpu/kenlm/archive/master.zip

run.sh

@@ -0,0 +1,38 @@
+python ~/xls-r-uzbek-cv8/run_speech_recognition_ctc.py \
+	--dataset_name="mozilla-foundation/common_voice_8_0" \
+	--model_name_or_path="facebook/wav2vec2-xls-r-300m" \
+	--dataset_config_name="uz" \
+	--output_dir="~/xls-r-uzbek-cv8" \
+	--train_split_name="train[:50%]" \
+	--eval_split_name="validation[50%:]" \
+	--overwrite_output_dir \
+	--num_train_epochs="100" \
+	--per_device_train_batch_size="32" \
+	--per_device_eval_batch_size="8" \
+	--gradient_accumulation_steps="4" \
+	--learning_rate="3e-5" \
+	--warmup_steps="500" \
+	--length_column_name="input_length" \
+	--evaluation_strategy="steps" \
+	--text_column_name="sentence" \
+	--eval_metrics wer cer \
+	--save_steps="500" \
+	--eval_steps="500" \
+	--logging_steps="100" \
+	--min_duration_in_seconds="0.2" \
+	--layerdrop="0.01" \
+	--activation_dropout="0.1" \
+	--save_total_limit="3" \
+	--freeze_feature_encoder \
+	--feat_proj_dropout="0.05" \
+	--mask_time_prob="0.50" \
+	--mask_time_length="10" \
+	--mask_feature_prob="0.15" \
+	--mask_feature_length="64" \
+	--gradient_checkpointing \
+	--use_auth_token \
+	--fp16 \
+	--group_by_length \
+	--do_train --do_eval \
+--push_to_hub
+#	--chars_to_ignore  \  # default to all punct

run_speech_recognition_ctc.py

@@ -0,0 +1,760 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+
+""" Fine-tuning a 🤗 Transformers CTC model for automatic speech recognition"""
+
+import functools
+import json
+import logging
+import os
+import re
+import string
+import sys
+import unidecode
+import warnings
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Union
+
+import datasets
+import numpy as np
+import torch
+from datasets import DatasetDict, load_dataset, load_metric
+
+import transformers
+from transformers import (
+    AutoConfig,
+    AutoFeatureExtractor,
+    AutoModelForCTC,
+    AutoProcessor,
+    AutoTokenizer,
+    HfArgumentParser,
+    Trainer,
+    TrainingArguments,
+    Wav2Vec2Processor,
+    set_seed,
+)
+from transformers.trainer_utils import get_last_checkpoint, is_main_process
+from transformers.utils import check_min_version
+from transformers.utils.versions import require_version
+
+
+# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
+check_min_version("4.16.0.dev0")
+
+require_version("datasets>=1.13.3", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
+
+
+logger = logging.getLogger(__name__)
+
+
+def list_field(default=None, metadata=None):
+    return field(default_factory=lambda: default, metadata=metadata)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
+    """
+
+    model_name_or_path: str = field(
+        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
+    )
+    tokenizer_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
+    )
+    cache_dir: Optional[str] = field(
+        default=None,
+        metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
+    )
+    freeze_feature_encoder: bool = field(
+        default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
+    )
+    attention_dropout: float = field(
+        default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
+    )
+    activation_dropout: float = field(
+        default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
+    )
+    feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
+    hidden_dropout: float = field(
+        default=0.0,
+        metadata={
+            "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
+        },
+    )
+    final_dropout: float = field(
+        default=0.0,
+        metadata={"help": "The dropout probability for the final projection layer."},
+    )
+    mask_time_prob: float = field(
+        default=0.05,
+        metadata={
+            "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
+            "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
+            "vectors will be masked along the time axis."
+        },
+    )
+    mask_time_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the time axis."},
+    )
+    mask_feature_prob: float = field(
+        default=0.0,
+        metadata={
+            "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
+            "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
+        },
+    )
+    mask_feature_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the feature axis."},
+    )
+    layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
+    ctc_loss_reduction: Optional[str] = field(
+        default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+
+    Using `HfArgumentParser` we can turn this class
+    into argparse arguments to be able to specify them on
+    the command line.
+    """
+
+    dataset_name: str = field(
+        metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: str = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    train_split_name: str = field(
+        default="train",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
+        },
+    )
+    eval_split_name: str = field(
+        default="validation",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'validation'"
+        },
+    )
+    audio_column_name: str = field(
+        default="audio",
+        metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
+    )
+    text_column_name: str = field(
+        default="text",
+        metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
+            "value if set."
+        },
+    )
+    chars_to_ignore: Optional[List[str]] = list_field(
+        default=None,
+        metadata={"help": "A list of characters to remove from the transcripts."},
+    )
+    eval_metrics: List[str] = list_field(
+        default=["wer"],
+        metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
+    )
+    max_duration_in_seconds: float = field(
+        default=20.0,
+        metadata={
+            "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
+        },
+    )
+    min_duration_in_seconds: float = field(
+        default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
+    )
+    preprocessing_only: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to only do data preprocessing and skip training. "
+            "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
+            "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
+            "so that the cached datasets can consequently be loaded in distributed training"
+        },
+    )
+    use_auth_token: bool = field(
+        default=False,
+        metadata={
+            "help": "If :obj:`True`, will use the token generated when running"
+            ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
+        },
+    )
+    unk_token: str = field(
+        default="[UNK]",
+        metadata={"help": "The unk token for the tokenizer"},
+    )
+    pad_token: str = field(
+        default="[PAD]",
+        metadata={"help": "The padding token for the tokenizer"},
+    )
+    word_delimiter_token: str = field(
+        default="|",
+        metadata={"help": "The word delimiter token for the tokenizer"},
+    )
+    phoneme_language: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The target language that should be used be"
+            " passed to the tokenizer for tokenization. Note that"
+            " this is only relevant if the model classifies the"
+            " input audio to a sequence of phoneme sequences."
+        },
+    )
+
+
+@dataclass
+class DataCollatorCTCWithPadding:
+    """
+    Data collator that will dynamically pad the inputs received.
+    Args:
+        processor (:class:`~transformers.AutoProcessor`)
+            The processor used for proccessing the data.
+        padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
+            Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
+            among:
+            * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
+              sequence if provided).
+            * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
+              maximum acceptable input length for the model if that argument is not provided.
+            * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
+              different lengths).
+        max_length (:obj:`int`, `optional`):
+            Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
+        max_length_labels (:obj:`int`, `optional`):
+            Maximum length of the ``labels`` returned list and optionally padding length (see above).
+        pad_to_multiple_of (:obj:`int`, `optional`):
+            If set will pad the sequence to a multiple of the provided value.
+            This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
+            7.5 (Volta).
+    """
+
+    processor: AutoProcessor
+    padding: Union[bool, str] = "longest"
+    pad_to_multiple_of: Optional[int] = None
+    pad_to_multiple_of_labels: Optional[int] = None
+
+    def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
+        # split inputs and labels since they have to be of different lenghts and need
+        # different padding methods
+        input_features = [{"input_values": feature["input_values"]} for feature in features]
+        label_features = [{"input_ids": feature["labels"]} for feature in features]
+
+        batch = self.processor.pad(
+            input_features,
+            padding=self.padding,
+            pad_to_multiple_of=self.pad_to_multiple_of,
+            return_tensors="pt",
+        )
+
+        with self.processor.as_target_processor():
+            labels_batch = self.processor.pad(
+                label_features,
+                padding=self.padding,
+                pad_to_multiple_of=self.pad_to_multiple_of_labels,
+                return_tensors="pt",
+            )
+
+        # replace padding with -100 to ignore loss correctly
+        labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
+
+        batch["labels"] = labels
+
+        return batch
+
+
+def create_vocabulary_from_data(
+    datasets: DatasetDict,
+    word_delimiter_token: Optional[str] = None,
+    unk_token: Optional[str] = None,
+    pad_token: Optional[str] = None,
+):
+    # Given training and test labels create vocabulary
+    def extract_all_chars(batch):
+        all_text = " ".join(batch["target_text"])
+        vocab = list(set(all_text))
+        return {"vocab": [vocab], "all_text": [all_text]}
+
+    vocabs = datasets.map(
+        extract_all_chars,
+        batched=True,
+        batch_size=-1,
+        keep_in_memory=True,
+        remove_columns=datasets["train"].column_names,
+    )
+
+    # take union of all unique characters in each dataset
+    vocab_set = functools.reduce(
+        lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
+    )
+
+    vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
+
+    # replace white space with delimiter token
+    if word_delimiter_token is not None:
+        vocab_dict[word_delimiter_token] = vocab_dict[" "]
+        del vocab_dict[" "]
+
+    # add unk and pad token
+    if unk_token is not None:
+        vocab_dict[unk_token] = len(vocab_dict)
+
+    if pad_token is not None:
+        vocab_dict[pad_token] = len(vocab_dict)
+
+    return vocab_dict
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    # Detecting last checkpoint.
+    last_checkpoint = None
+    if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
+        last_checkpoint = get_last_checkpoint(training_args.output_dir)
+        if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
+            raise ValueError(
+                f"Output directory ({training_args.output_dir}) already exists and is not empty. "
+                "Use --overwrite_output_dir to overcome."
+            )
+        elif last_checkpoint is not None:
+            logger.info(
+                f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
+                "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
+            )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        handlers=[logging.StreamHandler(sys.stdout)],
+    )
+    logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
+
+    # Log on each process the small summary:
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+        f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    )
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    if is_main_process(training_args.local_rank):
+        transformers.utils.logging.set_verbosity_info()
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # 1. First, let's load the dataset
+    raw_datasets = DatasetDict()
+
+    if training_args.do_train:
+        raw_datasets["train"] = load_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            split=data_args.train_split_name,
+            use_auth_token=data_args.use_auth_token,
+        )
+
+        if data_args.audio_column_name not in raw_datasets["train"].column_names:
+            raise ValueError(
+                f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
+                "Make sure to set `--audio_column_name` to the correct audio column - one of "
+                f"{', '.join(raw_datasets['train'].column_names)}."
+            )
+
+        if data_args.text_column_name not in raw_datasets["train"].column_names:
+            raise ValueError(
+                f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
+                "Make sure to set `--text_column_name` to the correct text column - one of "
+                f"{', '.join(raw_datasets['train'].column_names)}."
+            )
+
+        if data_args.max_train_samples is not None:
+            raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
+
+    if training_args.do_eval:
+        raw_datasets["eval"] = load_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            split=data_args.eval_split_name,
+            use_auth_token=data_args.use_auth_token,
+        )
+
+        if data_args.max_eval_samples is not None:
+            raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
+
+    # 2. We remove some special characters from the datasets
+    # that make training complicated and do not help in transcribing the speech
+    # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
+    # that could be easily picked up by the model
+    if data_args.chars_to_ignore is None:
+        chars_to_ignore_regex = f'[{re.escape(string.punctuation)}]'
+    else:
+        chars_to_ignore_regex = f'[{"".join(data_args.chars_to_ignore)}]'
+    print("chars_to_ignore", chars_to_ignore_regex)
+    text_column_name = data_args.text_column_name
+
+    def remove_special_characters(batch):
+        if chars_to_ignore_regex is not None:
+            batch["target_text"] = re.sub(
+                chars_to_ignore_regex, 
+                "", 
+                re.sub("['`´]", "’",   # elsewhere probably meant as glottal stop
+                       re.sub("([og])['`´]", "\g<1>‘",  # after o/g indicate modified char
+                              unidecode.unidecode(batch[text_column_name]).lower()
+                             )
+                      )
+            ) + " "
+        else:
+            batch["target_text"] = batch[text_column_name].lower() + " "
+        return batch
+
+    with training_args.main_process_first(desc="dataset map special characters removal"):
+        raw_datasets = raw_datasets.map(
+            remove_special_characters,
+            remove_columns=[text_column_name],
+            desc="remove special characters from datasets",
+        )
+
+    num_workers = data_args.preprocessing_num_workers
+
+    def is_transcript_in_length_range(text):
+        return 3 < len(text) < 200
+    
+    raw_datasets = raw_datasets.filter(
+        is_transcript_in_length_range,
+        num_proc=num_workers,
+        input_columns=["target_text"],
+    )
+    
+    # save special tokens for tokenizer
+    word_delimiter_token = data_args.word_delimiter_token
+    unk_token = data_args.unk_token
+    pad_token = data_args.pad_token
+
+    # 3. Next, let's load the config as we might need it to create
+    # the tokenizer
+    # load config
+    config = AutoConfig.from_pretrained(
+        model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
+    )
+
+    # 4. Next, if no tokenizer file is defined,
+    # we create the vocabulary of the model by extracting all unique characters from
+    # the training and evaluation datasets
+    # We need to make sure that only first rank saves vocabulary
+    # make sure all processes wait until vocab is created
+    tokenizer_name_or_path = model_args.tokenizer_name_or_path
+    tokenizer_kwargs = {}
+    if tokenizer_name_or_path is None:
+        # save vocab in training output dir
+        tokenizer_name_or_path = training_args.output_dir
+
+        vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
+
+        with training_args.main_process_first():
+            if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
+                os.remove(vocab_file)
+
+        with training_args.main_process_first(desc="dataset map vocabulary creation"):
+            if not os.path.isfile(vocab_file):
+                os.makedirs(tokenizer_name_or_path, exist_ok=True)
+                vocab_dict = create_vocabulary_from_data(
+                    raw_datasets,
+                    word_delimiter_token=word_delimiter_token,
+                    unk_token=unk_token,
+                    pad_token=pad_token,
+                )
+
+                # save vocab dict to be loaded into tokenizer
+                with open(vocab_file, "w") as file:
+                    json.dump(vocab_dict, file)
+
+        # if tokenizer has just been created
+        # it is defined by `tokenizer_class` if present in config else by `model_type`
+        tokenizer_kwargs = {
+            "config": config if config.tokenizer_class is not None else None,
+            "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
+            "unk_token": unk_token,
+            "pad_token": pad_token,
+            "word_delimiter_token": word_delimiter_token,
+        }
+
+    # 5. Now we can instantiate the feature extractor, tokenizer and model
+    # Note for distributed training, the .from_pretrained methods guarantee that only
+    # one local process can concurrently download model & vocab.
+
+    # load feature_extractor and tokenizer
+    tokenizer = AutoTokenizer.from_pretrained(
+        tokenizer_name_or_path,
+        use_auth_token=data_args.use_auth_token,
+        **tokenizer_kwargs,
+    )
+    feature_extractor = AutoFeatureExtractor.from_pretrained(
+        model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
+    )
+
+    # adapt config
+    config.update(
+        {
+            "feat_proj_dropout": model_args.feat_proj_dropout,
+            "attention_dropout": model_args.attention_dropout,
+            "hidden_dropout": model_args.hidden_dropout,
+            "final_dropout": model_args.final_dropout,
+            "mask_time_prob": model_args.mask_time_prob,
+            "mask_time_length": model_args.mask_time_length,
+            "mask_feature_prob": model_args.mask_feature_prob,
+            "mask_feature_length": model_args.mask_feature_length,
+            "gradient_checkpointing": training_args.gradient_checkpointing,
+            "layerdrop": model_args.layerdrop,
+            "ctc_loss_reduction": model_args.ctc_loss_reduction,
+            "ctc_zero_infinity": True,
+            "pad_token_id": tokenizer.pad_token_id,
+            "vocab_size": len(tokenizer),
+            "activation_dropout": model_args.activation_dropout,
+        }
+    )
+
+    # create model
+    model = AutoModelForCTC.from_pretrained(
+        model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        config=config,
+        use_auth_token=data_args.use_auth_token,
+    )
+
+    # freeze encoder
+    if model_args.freeze_feature_encoder:
+        model.freeze_feature_encoder()
+
+    # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
+    # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
+    # so that we just need to set the correct target sampling rate and normalize the input
+    # via the `feature_extractor`
+
+    # make sure that dataset decodes audio with correct sampling rate
+    dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
+    if dataset_sampling_rate != feature_extractor.sampling_rate:
+        raw_datasets = raw_datasets.cast_column(
+            data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
+        )
+
+    # derive max & min input length for sample rate & max duration
+    max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
+    min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
+    audio_column_name = data_args.audio_column_name
+
+    # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
+    phoneme_language = data_args.phoneme_language
+
+    # Preprocessing the datasets.
+    # We need to read the audio files as arrays and tokenize the targets.
+    def prepare_dataset(batch):
+        # load audio
+        sample = batch[audio_column_name]
+
+        inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
+        batch["input_values"] = inputs.input_values[0]
+        batch["input_length"] = len(batch["input_values"])
+
+        # encode targets
+        additional_kwargs = {}
+        if phoneme_language is not None:
+            additional_kwargs["phonemizer_lang"] = phoneme_language
+
+        batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
+        return batch
+
+    with training_args.main_process_first(desc="dataset map preprocessing"):
+        vectorized_datasets = raw_datasets.map(
+            prepare_dataset,
+            remove_columns=next(iter(raw_datasets.values())).column_names,
+            num_proc=num_workers,
+            desc="preprocess datasets",
+        )
+
+        def is_audio_in_length_range(length):
+            return length > min_input_length and length < max_input_length
+
+        # filter data that is shorter than min_input_length
+        vectorized_datasets = vectorized_datasets.filter(
+            is_audio_in_length_range,
+            num_proc=num_workers,
+            input_columns=["input_length"],
+        )
+
+    # 7. Next, we can prepare the training.
+    # Let's use word error rate (WER) as our evaluation metric,
+    # instantiate a data collator and the trainer
+
+    # Define evaluation metrics during training, *i.e.* word error rate, character error rate
+    eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
+
+    # for large datasets it is advised to run the preprocessing on a
+    # single machine first with ``args.preprocessing_only`` since there will mostly likely
+    # be a timeout when running the script in distributed mode.
+    # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
+    # cached dataset
+    if data_args.preprocessing_only:
+        logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
+        return
+
+    def compute_metrics(pred):
+        pred_logits = pred.predictions
+        pred_ids = np.argmax(pred_logits, axis=-1)
+
+        pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
+
+        pred_str = tokenizer.batch_decode(pred_ids)
+        # we do not want to group tokens when computing the metrics
+        label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
+
+        metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
+
+        return metrics
+
+    # Now save everything to be able to create a single processor later
+    if is_main_process(training_args.local_rank):
+        # save feature extractor, tokenizer and config
+        feature_extractor.save_pretrained(training_args.output_dir)
+        tokenizer.save_pretrained(training_args.output_dir)
+        config.save_pretrained(training_args.output_dir)
+
+    try:
+        processor = AutoProcessor.from_pretrained(training_args.output_dir)
+    except (OSError, KeyError):
+        warnings.warn(
+            "Loading a processor from a feature extractor config that does not"
+            " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
+            " attribute to your `preprocessor_config.json` file to suppress this warning: "
+            " `'processor_class': 'Wav2Vec2Processor'`",
+            FutureWarning,
+        )
+        processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
+
+    # Instantiate custom data collator
+    data_collator = DataCollatorCTCWithPadding(processor=processor)
+
+    # Initialize Trainer
+    trainer = Trainer(
+        model=model,
+        data_collator=data_collator,
+        args=training_args,
+        compute_metrics=compute_metrics,
+        train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
+        eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
+        tokenizer=feature_extractor,
+    )
+
+    # 8. Finally, we can start training
+
+    # Training
+    if training_args.do_train:
+
+        # use last checkpoint if exist
+        if last_checkpoint is not None:
+            checkpoint = last_checkpoint
+        elif os.path.isdir(model_args.model_name_or_path):
+            checkpoint = model_args.model_name_or_path
+        else:
+            checkpoint = None
+
+        train_result = trainer.train(resume_from_checkpoint=checkpoint)
+        trainer.save_model()
+
+        metrics = train_result.metrics
+        max_train_samples = (
+            data_args.max_train_samples
+            if data_args.max_train_samples is not None
+            else len(vectorized_datasets["train"])
+        )
+        metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
+
+        trainer.log_metrics("train", metrics)
+        trainer.save_metrics("train", metrics)
+        trainer.save_state()
+
+    # Evaluation
+    results = {}
+    if training_args.do_eval:
+        logger.info("*** Evaluate ***")
+        metrics = trainer.evaluate()
+        max_eval_samples = (
+            data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
+        )
+        metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
+
+        trainer.log_metrics("eval", metrics)
+        trainer.save_metrics("eval", metrics)
+
+    # Write model card and (optionally) push to hub
+    config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
+    kwargs = {
+        "finetuned_from": model_args.model_name_or_path,
+        "tasks": "speech-recognition",
+        "tags": ["automatic-speech-recognition", data_args.dataset_name],
+        "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
+        "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
+    }
+    if "common_voice" in data_args.dataset_name:
+        kwargs["language"] = config_name
+
+    if training_args.push_to_hub:
+        trainer.push_to_hub(**kwargs)
+    else:
+        trainer.create_model_card(**kwargs)
+
+    return results
+
+
+if __name__ == "__main__":
+    main()

special_tokens_map.json

@@ -0,0 +1,28 @@
+{
+  "additional_special_tokens": [
+    {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false
+    },
+    {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false
+    }
+  ],
+  "bos_token": "<s>",
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "[PAD]",
+  "unk_token": "[UNK]"
+}

tokenizer_config.json

@@ -0,0 +1,52 @@
+{
+  "added_tokens_decoder": {
+    "29": {
+      "content": "[UNK]",
+      "lstrip": true,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": false
+    },
+    "30": {
+      "content": "[PAD]",
+      "lstrip": true,
+      "normalized": false,
+      "rstrip": true,
+      "single_word": false,
+      "special": false
+    },
+    "31": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "32": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "additional_special_tokens": [
+    "<s>",
+    "</s>"
+  ],
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": true,
+  "do_lower_case": false,
+  "eos_token": "</s>",
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "[PAD]",
+  "processor_class": "Wav2Vec2ProcessorWithLM",
+  "replace_word_delimiter_char": " ",
+  "target_lang": null,
+  "tokenizer_class": "Wav2Vec2CTCTokenizer",
+  "unk_token": "[UNK]",
+  "word_delimiter_token": "|"
+}

train_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 100.0,
+    "train_loss": 0.9527478711016767,
+    "train_runtime": 101833.8069,
+    "train_samples": 19726,
+    "train_samples_per_second": 19.371,
+    "train_steps_per_second": 0.151
+}

trainer_state.json

@@ -0,0 +1,1249 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 99.99837925445705,
+  "global_step": 15400,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.65,
+      "learning_rate": 5.940000000000001e-06,
+      "loss": 11.9235,
+      "step": 100
+    },
+    {
+      "epoch": 1.3,
+      "learning_rate": 1.1940000000000001e-05,
+      "loss": 5.1849,
+      "step": 200
+    },
+    {
+      "epoch": 1.95,
+      "learning_rate": 1.794e-05,
+      "loss": 3.7405,
+      "step": 300
+    },
+    {
+      "epoch": 2.6,
+      "learning_rate": 2.394e-05,
+      "loss": 3.3052,
+      "step": 400
+    },
+    {
+      "epoch": 3.25,
+      "learning_rate": 2.994e-05,
+      "loss": 3.1401,
+      "step": 500
+    },
+    {
+      "epoch": 3.25,
+      "eval_cer": 1.0,
+      "eval_loss": 3.114637851715088,
+      "eval_runtime": 219.3333,
+      "eval_samples_per_second": 24.734,
+      "eval_steps_per_second": 3.096,
+      "eval_wer": 1.0,
+      "step": 500
+    },
+    {
+      "epoch": 3.89,
+      "learning_rate": 2.98006711409396e-05,
+      "loss": 3.0799,
+      "step": 600
+    },
+    {
+      "epoch": 4.54,
+      "learning_rate": 2.9599328859060405e-05,
+      "loss": 3.0534,
+      "step": 700
+    },
+    {
+      "epoch": 5.19,
+      "learning_rate": 2.9397986577181207e-05,
+      "loss": 2.994,
+      "step": 800
+    },
+    {
+      "epoch": 5.84,
+      "learning_rate": 2.9196644295302013e-05,
+      "loss": 2.9428,
+      "step": 900
+    },
+    {
+      "epoch": 6.49,
+      "learning_rate": 2.8995302013422818e-05,
+      "loss": 2.7484,
+      "step": 1000
+    },
+    {
+      "epoch": 6.49,
+      "eval_cer": 0.706927100586246,
+      "eval_loss": 2.284235954284668,
+      "eval_runtime": 222.6226,
+      "eval_samples_per_second": 24.369,
+      "eval_steps_per_second": 3.05,
+      "eval_wer": 1.0065119583234667,
+      "step": 1000
+    },
+    {
+      "epoch": 7.14,
+      "learning_rate": 2.8793959731543624e-05,
+      "loss": 2.0772,
+      "step": 1100
+    },
+    {
+      "epoch": 7.79,
+      "learning_rate": 2.859261744966443e-05,
+      "loss": 1.4967,
+      "step": 1200
+    },
+    {
+      "epoch": 8.44,
+      "learning_rate": 2.8391275167785235e-05,
+      "loss": 1.2731,
+      "step": 1300
+    },
+    {
+      "epoch": 9.09,
+      "learning_rate": 2.818993288590604e-05,
+      "loss": 1.1742,
+      "step": 1400
+    },
+    {
+      "epoch": 9.74,
+      "learning_rate": 2.7988590604026846e-05,
+      "loss": 1.0899,
+      "step": 1500
+    },
+    {
+      "epoch": 9.74,
+      "eval_cer": 0.13506382454221202,
+      "eval_loss": 0.5414122343063354,
+      "eval_runtime": 217.7702,
+      "eval_samples_per_second": 24.912,
+      "eval_steps_per_second": 3.118,
+      "eval_wer": 0.6124792801326071,
+      "step": 1500
+    },
+    {
+      "epoch": 10.39,
+      "learning_rate": 2.778724832214765e-05,
+      "loss": 1.0544,
+      "step": 1600
+    },
+    {
+      "epoch": 11.04,
+      "learning_rate": 2.7585906040268457e-05,
+      "loss": 1.0284,
+      "step": 1700
+    },
+    {
+      "epoch": 11.69,
+      "learning_rate": 2.7384563758389263e-05,
+      "loss": 0.9865,
+      "step": 1800
+    },
+    {
+      "epoch": 12.34,
+      "learning_rate": 2.7183221476510065e-05,
+      "loss": 0.9705,
+      "step": 1900
+    },
+    {
+      "epoch": 12.99,
+      "learning_rate": 2.698187919463087e-05,
+      "loss": 0.9465,
+      "step": 2000
+    },
+    {
+      "epoch": 12.99,
+      "eval_cer": 0.12229891609980818,
+      "eval_loss": 0.4565887749195099,
+      "eval_runtime": 219.4838,
+      "eval_samples_per_second": 24.717,
+      "eval_steps_per_second": 3.094,
+      "eval_wer": 0.5634619938432394,
+      "step": 2000
+    },
+    {
+      "epoch": 13.64,
+      "learning_rate": 2.6780536912751676e-05,
+      "loss": 0.9395,
+      "step": 2100
+    },
+    {
+      "epoch": 14.29,
+      "learning_rate": 2.6579194630872482e-05,
+      "loss": 0.9216,
+      "step": 2200
+    },
+    {
+      "epoch": 14.93,
+      "learning_rate": 2.6377852348993287e-05,
+      "loss": 0.9011,
+      "step": 2300
+    },
+    {
+      "epoch": 15.58,
+      "learning_rate": 2.6176510067114093e-05,
+      "loss": 0.9011,
+      "step": 2400
+    },
+    {
+      "epoch": 16.23,
+      "learning_rate": 2.59751677852349e-05,
+      "loss": 0.8771,
+      "step": 2500
+    },
+    {
+      "epoch": 16.23,
+      "eval_cer": 0.11608208983968754,
+      "eval_loss": 0.42124298214912415,
+      "eval_runtime": 218.6532,
+      "eval_samples_per_second": 24.811,
+      "eval_steps_per_second": 3.105,
+      "eval_wer": 0.5365557660430973,
+      "step": 2500
+    },
+    {
+      "epoch": 16.88,
+      "learning_rate": 2.5773825503355704e-05,
+      "loss": 0.8719,
+      "step": 2600
+    },
+    {
+      "epoch": 17.53,
+      "learning_rate": 2.557248322147651e-05,
+      "loss": 0.8616,
+      "step": 2700
+    },
+    {
+      "epoch": 18.18,
+      "learning_rate": 2.5371140939597315e-05,
+      "loss": 0.8573,
+      "step": 2800
+    },
+    {
+      "epoch": 18.83,
+      "learning_rate": 2.5169798657718124e-05,
+      "loss": 0.8467,
+      "step": 2900
+    },
+    {
+      "epoch": 19.48,
+      "learning_rate": 2.496845637583893e-05,
+      "loss": 0.8346,
+      "step": 3000
+    },
+    {
+      "epoch": 19.48,
+      "eval_cer": 0.11022733400611667,
+      "eval_loss": 0.3994467258453369,
+      "eval_runtime": 213.3023,
+      "eval_samples_per_second": 25.433,
+      "eval_steps_per_second": 3.183,
+      "eval_wer": 0.5143855079327492,
+      "step": 3000
+    },
+    {
+      "epoch": 20.13,
+      "learning_rate": 2.4767114093959732e-05,
+      "loss": 0.8386,
+      "step": 3100
+    },
+    {
+      "epoch": 20.78,
+      "learning_rate": 2.4565771812080538e-05,
+      "loss": 0.8283,
+      "step": 3200
+    },
+    {
+      "epoch": 21.43,
+      "learning_rate": 2.4364429530201343e-05,
+      "loss": 0.8198,
+      "step": 3300
+    },
+    {
+      "epoch": 22.08,
+      "learning_rate": 2.416308724832215e-05,
+      "loss": 0.8164,
+      "step": 3400
+    },
+    {
+      "epoch": 22.73,
+      "learning_rate": 2.3961744966442955e-05,
+      "loss": 0.8127,
+      "step": 3500
+    },
+    {
+      "epoch": 22.73,
+      "eval_cer": 0.10510827446479058,
+      "eval_loss": 0.3818908929824829,
+      "eval_runtime": 218.4137,
+      "eval_samples_per_second": 24.838,
+      "eval_steps_per_second": 3.109,
+      "eval_wer": 0.49437603599336966,
+      "step": 3500
+    },
+    {
+      "epoch": 23.38,
+      "learning_rate": 2.376040268456376e-05,
+      "loss": 0.8081,
+      "step": 3600
+    },
+    {
+      "epoch": 24.03,
+      "learning_rate": 2.3559060402684566e-05,
+      "loss": 0.7988,
+      "step": 3700
+    },
+    {
+      "epoch": 24.67,
+      "learning_rate": 2.335771812080537e-05,
+      "loss": 0.7903,
+      "step": 3800
+    },
+    {
+      "epoch": 25.32,
+      "learning_rate": 2.3156375838926177e-05,
+      "loss": 0.788,
+      "step": 3900
+    },
+    {
+      "epoch": 25.97,
+      "learning_rate": 2.2955033557046982e-05,
+      "loss": 0.7833,
+      "step": 4000
+    },
+    {
+      "epoch": 25.97,
+      "eval_cer": 0.10109853708140422,
+      "eval_loss": 0.3704679310321808,
+      "eval_runtime": 216.2288,
+      "eval_samples_per_second": 25.089,
+      "eval_steps_per_second": 3.14,
+      "eval_wer": 0.4797537295761307,
+      "step": 4000
+    },
+    {
+      "epoch": 26.62,
+      "learning_rate": 2.2753691275167788e-05,
+      "loss": 0.7724,
+      "step": 4100
+    },
+    {
+      "epoch": 27.27,
+      "learning_rate": 2.255234899328859e-05,
+      "loss": 0.7695,
+      "step": 4200
+    },
+    {
+      "epoch": 27.92,
+      "learning_rate": 2.2351006711409396e-05,
+      "loss": 0.7653,
+      "step": 4300
+    },
+    {
+      "epoch": 28.57,
+      "learning_rate": 2.2151677852348994e-05,
+      "loss": 0.7728,
+      "step": 4400
+    },
+    {
+      "epoch": 29.22,
+      "learning_rate": 2.19503355704698e-05,
+      "loss": 0.7603,
+      "step": 4500
+    },
+    {
+      "epoch": 29.22,
+      "eval_cer": 0.0992381113790261,
+      "eval_loss": 0.36611661314964294,
+      "eval_runtime": 217.7866,
+      "eval_samples_per_second": 24.91,
+      "eval_steps_per_second": 3.118,
+      "eval_wer": 0.4704001894387876,
+      "step": 4500
+    },
+    {
+      "epoch": 29.87,
+      "learning_rate": 2.1748993288590605e-05,
+      "loss": 0.7563,
+      "step": 4600
+    },
+    {
+      "epoch": 30.52,
+      "learning_rate": 2.154765100671141e-05,
+      "loss": 0.7618,
+      "step": 4700
+    },
+    {
+      "epoch": 31.17,
+      "learning_rate": 2.1346308724832217e-05,
+      "loss": 0.7467,
+      "step": 4800
+    },
+    {
+      "epoch": 31.82,
+      "learning_rate": 2.114697986577181e-05,
+      "loss": 0.7514,
+      "step": 4900
+    },
+    {
+      "epoch": 32.47,
+      "learning_rate": 2.0945637583892617e-05,
+      "loss": 0.7424,
+      "step": 5000
+    },
+    {
+      "epoch": 32.47,
+      "eval_cer": 0.09569444337449638,
+      "eval_loss": 0.3528956174850464,
+      "eval_runtime": 218.9072,
+      "eval_samples_per_second": 24.782,
+      "eval_steps_per_second": 3.102,
+      "eval_wer": 0.4577314705185887,
+      "step": 5000
+    },
+    {
+      "epoch": 33.12,
+      "learning_rate": 2.0744295302013423e-05,
+      "loss": 0.748,
+      "step": 5100
+    },
+    {
+      "epoch": 33.76,
+      "learning_rate": 2.054295302013423e-05,
+      "loss": 0.7357,
+      "step": 5200
+    },
+    {
+      "epoch": 34.41,
+      "learning_rate": 2.0341610738255034e-05,
+      "loss": 0.7357,
+      "step": 5300
+    },
+    {
+      "epoch": 35.06,
+      "learning_rate": 2.014026845637584e-05,
+      "loss": 0.735,
+      "step": 5400
+    },
+    {
+      "epoch": 35.71,
+      "learning_rate": 1.9938926174496645e-05,
+      "loss": 0.7251,
+      "step": 5500
+    },
+    {
+      "epoch": 35.71,
+      "eval_cer": 0.09283254627953377,
+      "eval_loss": 0.34103500843048096,
+      "eval_runtime": 218.8353,
+      "eval_samples_per_second": 24.79,
+      "eval_steps_per_second": 3.103,
+      "eval_wer": 0.4472827373904807,
+      "step": 5500
+    },
+    {
+      "epoch": 36.36,
+      "learning_rate": 1.973758389261745e-05,
+      "loss": 0.7306,
+      "step": 5600
+    },
+    {
+      "epoch": 37.01,
+      "learning_rate": 1.9536241610738256e-05,
+      "loss": 0.7185,
+      "step": 5700
+    },
+    {
+      "epoch": 37.66,
+      "learning_rate": 1.9334899328859062e-05,
+      "loss": 0.7135,
+      "step": 5800
+    },
+    {
+      "epoch": 38.31,
+      "learning_rate": 1.9133557046979864e-05,
+      "loss": 0.726,
+      "step": 5900
+    },
+    {
+      "epoch": 38.96,
+      "learning_rate": 1.893221476510067e-05,
+      "loss": 0.7106,
+      "step": 6000
+    },
+    {
+      "epoch": 38.96,
+      "eval_cer": 0.09190425933486893,
+      "eval_loss": 0.3401394486427307,
+      "eval_runtime": 218.6772,
+      "eval_samples_per_second": 24.808,
+      "eval_steps_per_second": 3.105,
+      "eval_wer": 0.4427539663746152,
+      "step": 6000
+    },
+    {
+      "epoch": 39.61,
+      "learning_rate": 1.8730872483221475e-05,
+      "loss": 0.7157,
+      "step": 6100
+    },
+    {
+      "epoch": 40.26,
+      "learning_rate": 1.852953020134228e-05,
+      "loss": 0.7055,
+      "step": 6200
+    },
+    {
+      "epoch": 40.91,
+      "learning_rate": 1.8328187919463086e-05,
+      "loss": 0.7108,
+      "step": 6300
+    },
+    {
+      "epoch": 41.56,
+      "learning_rate": 1.8126845637583892e-05,
+      "loss": 0.705,
+      "step": 6400
+    },
+    {
+      "epoch": 42.21,
+      "learning_rate": 1.7925503355704698e-05,
+      "loss": 0.7027,
+      "step": 6500
+    },
+    {
+      "epoch": 42.21,
+      "eval_cer": 0.09045597762866982,
+      "eval_loss": 0.3354834318161011,
+      "eval_runtime": 219.2927,
+      "eval_samples_per_second": 24.739,
+      "eval_steps_per_second": 3.096,
+      "eval_wer": 0.4352652143026285,
+      "step": 6500
+    },
+    {
+      "epoch": 42.86,
+      "learning_rate": 1.7724161073825503e-05,
+      "loss": 0.7103,
+      "step": 6600
+    },
+    {
+      "epoch": 43.51,
+      "learning_rate": 1.752281879194631e-05,
+      "loss": 0.7011,
+      "step": 6700
+    },
+    {
+      "epoch": 44.16,
+      "learning_rate": 1.7321476510067114e-05,
+      "loss": 0.7004,
+      "step": 6800
+    },
+    {
+      "epoch": 44.8,
+      "learning_rate": 1.712013422818792e-05,
+      "loss": 0.702,
+      "step": 6900
+    },
+    {
+      "epoch": 45.45,
+      "learning_rate": 1.6918791946308722e-05,
+      "loss": 0.6927,
+      "step": 7000
+    },
+    {
+      "epoch": 45.45,
+      "eval_cer": 0.08853392291751727,
+      "eval_loss": 0.33077025413513184,
+      "eval_runtime": 215.9851,
+      "eval_samples_per_second": 25.117,
+      "eval_steps_per_second": 3.144,
+      "eval_wer": 0.4296412502959981,
+      "step": 7000
+    },
+    {
+      "epoch": 46.1,
+      "learning_rate": 1.6717449664429528e-05,
+      "loss": 0.691,
+      "step": 7100
+    },
+    {
+      "epoch": 46.75,
+      "learning_rate": 1.6516107382550333e-05,
+      "loss": 0.6833,
+      "step": 7200
+    },
+    {
+      "epoch": 47.4,
+      "learning_rate": 1.631476510067114e-05,
+      "loss": 0.692,
+      "step": 7300
+    },
+    {
+      "epoch": 48.05,
+      "learning_rate": 1.6113422818791948e-05,
+      "loss": 0.6871,
+      "step": 7400
+    },
+    {
+      "epoch": 48.7,
+      "learning_rate": 1.5914093959731546e-05,
+      "loss": 0.6828,
+      "step": 7500
+    },
+    {
+      "epoch": 48.7,
+      "eval_cer": 0.08625364959286336,
+      "eval_loss": 0.324627548456192,
+      "eval_runtime": 221.033,
+      "eval_samples_per_second": 24.544,
+      "eval_steps_per_second": 3.072,
+      "eval_wer": 0.42043570921146106,
+      "step": 7500
+    },
+    {
+      "epoch": 49.35,
+      "learning_rate": 1.5712751677852352e-05,
+      "loss": 0.6789,
+      "step": 7600
+    },
+    {
+      "epoch": 50.0,
+      "learning_rate": 1.5511409395973158e-05,
+      "loss": 0.6811,
+      "step": 7700
+    },
+    {
+      "epoch": 50.65,
+      "learning_rate": 1.531006711409396e-05,
+      "loss": 0.683,
+      "step": 7800
+    },
+    {
+      "epoch": 51.3,
+      "learning_rate": 1.5108724832214764e-05,
+      "loss": 0.6765,
+      "step": 7900
+    },
+    {
+      "epoch": 51.95,
+      "learning_rate": 1.4907382550335571e-05,
+      "loss": 0.6706,
+      "step": 8000
+    },
+    {
+      "epoch": 51.95,
+      "eval_cer": 0.08681216248488163,
+      "eval_loss": 0.32503771781921387,
+      "eval_runtime": 215.1259,
+      "eval_samples_per_second": 25.218,
+      "eval_steps_per_second": 3.156,
+      "eval_wer": 0.42327729102533745,
+      "step": 8000
+    },
+    {
+      "epoch": 52.6,
+      "learning_rate": 1.4706040268456375e-05,
+      "loss": 0.6777,
+      "step": 8100
+    },
+    {
+      "epoch": 53.25,
+      "learning_rate": 1.450469798657718e-05,
+      "loss": 0.6754,
+      "step": 8200
+    },
+    {
+      "epoch": 53.89,
+      "learning_rate": 1.4303355704697986e-05,
+      "loss": 0.6675,
+      "step": 8300
+    },
+    {
+      "epoch": 54.54,
+      "learning_rate": 1.4102013422818792e-05,
+      "loss": 0.6627,
+      "step": 8400
+    },
+    {
+      "epoch": 55.19,
+      "learning_rate": 1.3900671140939599e-05,
+      "loss": 0.6629,
+      "step": 8500
+    },
+    {
+      "epoch": 55.19,
+      "eval_cer": 0.0849055150259227,
+      "eval_loss": 0.3263927102088928,
+      "eval_runtime": 213.0987,
+      "eval_samples_per_second": 25.458,
+      "eval_steps_per_second": 3.186,
+      "eval_wer": 0.4158773383850343,
+      "step": 8500
+    },
+    {
+      "epoch": 55.84,
+      "learning_rate": 1.3699328859060405e-05,
+      "loss": 0.6632,
+      "step": 8600
+    },
+    {
+      "epoch": 56.49,
+      "learning_rate": 1.3497986577181208e-05,
+      "loss": 0.6558,
+      "step": 8700
+    },
+    {
+      "epoch": 57.14,
+      "learning_rate": 1.3296644295302014e-05,
+      "loss": 0.6691,
+      "step": 8800
+    },
+    {
+      "epoch": 57.79,
+      "learning_rate": 1.309530201342282e-05,
+      "loss": 0.6633,
+      "step": 8900
+    },
+    {
+      "epoch": 58.44,
+      "learning_rate": 1.2893959731543625e-05,
+      "loss": 0.6556,
+      "step": 9000
+    },
+    {
+      "epoch": 58.44,
+      "eval_cer": 0.08353812139374003,
+      "eval_loss": 0.3212815821170807,
+      "eval_runtime": 220.228,
+      "eval_samples_per_second": 24.634,
+      "eval_steps_per_second": 3.083,
+      "eval_wer": 0.40995737627279183,
+      "step": 9000
+    },
+    {
+      "epoch": 59.09,
+      "learning_rate": 1.269261744966443e-05,
+      "loss": 0.6584,
+      "step": 9100
+    },
+    {
+      "epoch": 59.74,
+      "learning_rate": 1.2491275167785236e-05,
+      "loss": 0.6537,
+      "step": 9200
+    },
+    {
+      "epoch": 60.39,
+      "learning_rate": 1.228993288590604e-05,
+      "loss": 0.6633,
+      "step": 9300
+    },
+    {
+      "epoch": 61.04,
+      "learning_rate": 1.2088590604026846e-05,
+      "loss": 0.6474,
+      "step": 9400
+    },
+    {
+      "epoch": 61.69,
+      "learning_rate": 1.1887248322147651e-05,
+      "loss": 0.6484,
+      "step": 9500
+    },
+    {
+      "epoch": 61.69,
+      "eval_cer": 0.0836613794112889,
+      "eval_loss": 0.31816166639328003,
+      "eval_runtime": 216.5133,
+      "eval_samples_per_second": 25.056,
+      "eval_steps_per_second": 3.136,
+      "eval_wer": 0.41241416054937247,
+      "step": 9500
+    },
+    {
+      "epoch": 62.34,
+      "learning_rate": 1.1687919463087248e-05,
+      "loss": 0.6547,
+      "step": 9600
+    },
+    {
+      "epoch": 62.99,
+      "learning_rate": 1.1486577181208054e-05,
+      "loss": 0.6481,
+      "step": 9700
+    },
+    {
+      "epoch": 63.64,
+      "learning_rate": 1.128523489932886e-05,
+      "loss": 0.648,
+      "step": 9800
+    },
+    {
+      "epoch": 64.29,
+      "learning_rate": 1.1083892617449665e-05,
+      "loss": 0.6471,
+      "step": 9900
+    },
+    {
+      "epoch": 64.93,
+      "learning_rate": 1.088255033557047e-05,
+      "loss": 0.6407,
+      "step": 10000
+    },
+    {
+      "epoch": 64.93,
+      "eval_cer": 0.08248272461847792,
+      "eval_loss": 0.3171332776546478,
+      "eval_runtime": 217.1179,
+      "eval_samples_per_second": 24.986,
+      "eval_steps_per_second": 3.127,
+      "eval_wer": 0.4050142079090694,
+      "step": 10000
+    },
+    {
+      "epoch": 65.58,
+      "learning_rate": 1.0681208053691274e-05,
+      "loss": 0.6446,
+      "step": 10100
+    },
+    {
+      "epoch": 66.23,
+      "learning_rate": 1.047986577181208e-05,
+      "loss": 0.6383,
+      "step": 10200
+    },
+    {
+      "epoch": 66.88,
+      "learning_rate": 1.0278523489932886e-05,
+      "loss": 0.6413,
+      "step": 10300
+    },
+    {
+      "epoch": 67.53,
+      "learning_rate": 1.0077181208053691e-05,
+      "loss": 0.6494,
+      "step": 10400
+    },
+    {
+      "epoch": 68.18,
+      "learning_rate": 9.875838926174497e-06,
+      "loss": 0.6375,
+      "step": 10500
+    },
+    {
+      "epoch": 68.18,
+      "eval_cer": 0.0822362085833802,
+      "eval_loss": 0.31498104333877563,
+      "eval_runtime": 218.6829,
+      "eval_samples_per_second": 24.808,
+      "eval_steps_per_second": 3.105,
+      "eval_wer": 0.4038598152971821,
+      "step": 10500
+    },
+    {
+      "epoch": 68.83,
+      "learning_rate": 9.6744966442953e-06,
+      "loss": 0.6359,
+      "step": 10600
+    },
+    {
+      "epoch": 69.48,
+      "learning_rate": 9.473154362416108e-06,
+      "loss": 0.638,
+      "step": 10700
+    },
+    {
+      "epoch": 70.13,
+      "learning_rate": 9.271812080536914e-06,
+      "loss": 0.6405,
+      "step": 10800
+    },
+    {
+      "epoch": 70.78,
+      "learning_rate": 9.070469798657719e-06,
+      "loss": 0.6388,
+      "step": 10900
+    },
+    {
+      "epoch": 71.43,
+      "learning_rate": 8.869127516778525e-06,
+      "loss": 0.6363,
+      "step": 11000
+    },
+    {
+      "epoch": 71.43,
+      "eval_cer": 0.08095355483826237,
+      "eval_loss": 0.3129253089427948,
+      "eval_runtime": 218.9143,
+      "eval_samples_per_second": 24.781,
+      "eval_steps_per_second": 3.102,
+      "eval_wer": 0.3991238456073881,
+      "step": 11000
+    },
+    {
+      "epoch": 72.08,
+      "learning_rate": 8.66778523489933e-06,
+      "loss": 0.6369,
+      "step": 11100
+    },
+    {
+      "epoch": 72.73,
+      "learning_rate": 8.466442953020134e-06,
+      "loss": 0.635,
+      "step": 11200
+    },
+    {
+      "epoch": 73.38,
+      "learning_rate": 8.26510067114094e-06,
+      "loss": 0.6337,
+      "step": 11300
+    },
+    {
+      "epoch": 74.03,
+      "learning_rate": 8.063758389261745e-06,
+      "loss": 0.6308,
+      "step": 11400
+    },
+    {
+      "epoch": 74.67,
+      "learning_rate": 7.862416107382551e-06,
+      "loss": 0.6307,
+      "step": 11500
+    },
+    {
+      "epoch": 74.67,
+      "eval_cer": 0.08074170512060026,
+      "eval_loss": 0.3114279508590698,
+      "eval_runtime": 219.6542,
+      "eval_samples_per_second": 24.698,
+      "eval_steps_per_second": 3.091,
+      "eval_wer": 0.3986206488278475,
+      "step": 11500
+    },
+    {
+      "epoch": 75.32,
+      "learning_rate": 7.661073825503357e-06,
+      "loss": 0.6335,
+      "step": 11600
+    },
+    {
+      "epoch": 75.97,
+      "learning_rate": 7.459731543624161e-06,
+      "loss": 0.628,
+      "step": 11700
+    },
+    {
+      "epoch": 76.62,
+      "learning_rate": 7.260402684563759e-06,
+      "loss": 0.6324,
+      "step": 11800
+    },
+    {
+      "epoch": 77.27,
+      "learning_rate": 7.059060402684564e-06,
+      "loss": 0.6317,
+      "step": 11900
+    },
+    {
+      "epoch": 77.92,
+      "learning_rate": 6.857718120805369e-06,
+      "loss": 0.6232,
+      "step": 12000
+    },
+    {
+      "epoch": 77.92,
+      "eval_cer": 0.07899683380967422,
+      "eval_loss": 0.31030353903770447,
+      "eval_runtime": 220.4971,
+      "eval_samples_per_second": 24.603,
+      "eval_steps_per_second": 3.079,
+      "eval_wer": 0.3895335069855553,
+      "step": 12000
+    },
+    {
+      "epoch": 78.57,
+      "learning_rate": 6.656375838926175e-06,
+      "loss": 0.6295,
+      "step": 12100
+    },
+    {
+      "epoch": 79.22,
+      "learning_rate": 6.4550335570469795e-06,
+      "loss": 0.6234,
+      "step": 12200
+    },
+    {
+      "epoch": 79.87,
+      "learning_rate": 6.255704697986578e-06,
+      "loss": 0.6172,
+      "step": 12300
+    },
+    {
+      "epoch": 80.52,
+      "learning_rate": 6.054362416107383e-06,
+      "loss": 0.6203,
+      "step": 12400
+    },
+    {
+      "epoch": 81.17,
+      "learning_rate": 5.853020134228188e-06,
+      "loss": 0.6216,
+      "step": 12500
+    },
+    {
+      "epoch": 81.17,
+      "eval_cer": 0.0789506120530934,
+      "eval_loss": 0.30863967537879944,
+      "eval_runtime": 218.4142,
+      "eval_samples_per_second": 24.838,
+      "eval_steps_per_second": 3.109,
+      "eval_wer": 0.3891191096376983,
+      "step": 12500
+    },
+    {
+      "epoch": 81.82,
+      "learning_rate": 5.651677852348994e-06,
+      "loss": 0.6203,
+      "step": 12600
+    },
+    {
+      "epoch": 82.47,
+      "learning_rate": 5.4503355704697986e-06,
+      "loss": 0.6209,
+      "step": 12700
+    },
+    {
+      "epoch": 83.12,
+      "learning_rate": 5.248993288590604e-06,
+      "loss": 0.6257,
+      "step": 12800
+    },
+    {
+      "epoch": 83.76,
+      "learning_rate": 5.04765100671141e-06,
+      "loss": 0.6245,
+      "step": 12900
+    },
+    {
+      "epoch": 84.41,
+      "learning_rate": 4.8463087248322145e-06,
+      "loss": 0.6174,
+      "step": 13000
+    },
+    {
+      "epoch": 84.41,
+      "eval_cer": 0.07851150536557558,
+      "eval_loss": 0.3082079291343689,
+      "eval_runtime": 215.0269,
+      "eval_samples_per_second": 25.229,
+      "eval_steps_per_second": 3.158,
+      "eval_wer": 0.3880535164574947,
+      "step": 13000
+    },
+    {
+      "epoch": 85.06,
+      "learning_rate": 4.64496644295302e-06,
+      "loss": 0.6222,
+      "step": 13100
+    },
+    {
+      "epoch": 85.71,
+      "learning_rate": 4.443624161073826e-06,
+      "loss": 0.6113,
+      "step": 13200
+    },
+    {
+      "epoch": 86.36,
+      "learning_rate": 4.2422818791946304e-06,
+      "loss": 0.6238,
+      "step": 13300
+    },
+    {
+      "epoch": 87.01,
+      "learning_rate": 4.040939597315437e-06,
+      "loss": 0.618,
+      "step": 13400
+    },
+    {
+      "epoch": 87.66,
+      "learning_rate": 3.8395973154362425e-06,
+      "loss": 0.6196,
+      "step": 13500
+    },
+    {
+      "epoch": 87.66,
+      "eval_cer": 0.07821876757389704,
+      "eval_loss": 0.30590009689331055,
+      "eval_runtime": 213.2382,
+      "eval_samples_per_second": 25.441,
+      "eval_steps_per_second": 3.184,
+      "eval_wer": 0.3874911200568316,
+      "step": 13500
+    },
+    {
+      "epoch": 88.31,
+      "learning_rate": 3.6382550335570468e-06,
+      "loss": 0.6174,
+      "step": 13600
+    },
+    {
+      "epoch": 88.96,
+      "learning_rate": 3.4369127516778524e-06,
+      "loss": 0.6128,
+      "step": 13700
+    },
+    {
+      "epoch": 89.61,
+      "learning_rate": 3.235570469798658e-06,
+      "loss": 0.6246,
+      "step": 13800
+    },
+    {
+      "epoch": 90.26,
+      "learning_rate": 3.034228187919463e-06,
+      "loss": 0.6097,
+      "step": 13900
+    },
+    {
+      "epoch": 90.91,
+      "learning_rate": 2.8328859060402687e-06,
+      "loss": 0.6174,
+      "step": 14000
+    },
+    {
+      "epoch": 90.91,
+      "eval_cer": 0.07799151060404132,
+      "eval_loss": 0.30842480063438416,
+      "eval_runtime": 212.6251,
+      "eval_samples_per_second": 25.514,
+      "eval_steps_per_second": 3.193,
+      "eval_wer": 0.3862479280132607,
+      "step": 14000
+    },
+    {
+      "epoch": 91.56,
+      "learning_rate": 2.631543624161074e-06,
+      "loss": 0.6194,
+      "step": 14100
+    },
+    {
+      "epoch": 92.21,
+      "learning_rate": 2.430201342281879e-06,
+      "loss": 0.6167,
+      "step": 14200
+    },
+    {
+      "epoch": 92.86,
+      "learning_rate": 2.2288590604026842e-06,
+      "loss": 0.614,
+      "step": 14300
+    },
+    {
+      "epoch": 93.51,
+      "learning_rate": 2.0275167785234902e-06,
+      "loss": 0.615,
+      "step": 14400
+    },
+    {
+      "epoch": 94.16,
+      "learning_rate": 1.8261744966442954e-06,
+      "loss": 0.6169,
+      "step": 14500
+    },
+    {
+      "epoch": 94.16,
+      "eval_cer": 0.07787595621258926,
+      "eval_loss": 0.30701127648353577,
+      "eval_runtime": 215.1709,
+      "eval_samples_per_second": 25.213,
+      "eval_steps_per_second": 3.156,
+      "eval_wer": 0.3859519299076486,
+      "step": 14500
+    },
+    {
+      "epoch": 94.8,
+      "learning_rate": 1.6248322147651008e-06,
+      "loss": 0.6123,
+      "step": 14600
+    },
+    {
+      "epoch": 95.45,
+      "learning_rate": 1.423489932885906e-06,
+      "loss": 0.6133,
+      "step": 14700
+    },
+    {
+      "epoch": 96.1,
+      "learning_rate": 1.2221476510067115e-06,
+      "loss": 0.6068,
+      "step": 14800
+    },
+    {
+      "epoch": 96.75,
+      "learning_rate": 1.020805369127517e-06,
+      "loss": 0.6135,
+      "step": 14900
+    },
+    {
+      "epoch": 97.4,
+      "learning_rate": 8.194630872483221e-07,
+      "loss": 0.6166,
+      "step": 15000
+    },
+    {
+      "epoch": 97.4,
+      "eval_cer": 0.07776810544723402,
+      "eval_loss": 0.30662447214126587,
+      "eval_runtime": 214.9469,
+      "eval_samples_per_second": 25.239,
+      "eval_steps_per_second": 3.159,
+      "eval_wer": 0.3855079327492304,
+      "step": 15000
+    },
+    {
+      "epoch": 98.05,
+      "learning_rate": 6.181208053691276e-07,
+      "loss": 0.6189,
+      "step": 15100
+    },
+    {
+      "epoch": 98.7,
+      "learning_rate": 4.167785234899329e-07,
+      "loss": 0.6085,
+      "step": 15200
+    },
+    {
+      "epoch": 99.35,
+      "learning_rate": 2.1543624161073826e-07,
+      "loss": 0.6135,
+      "step": 15300
+    },
+    {
+      "epoch": 100.0,
+      "learning_rate": 1.4093959731543625e-08,
+      "loss": 0.6047,
+      "step": 15400
+    },
+    {
+      "epoch": 100.0,
+      "step": 15400,
+      "total_flos": 2.652008062738907e+20,
+      "train_loss": 0.9527478711016767,
+      "train_runtime": 101833.8069,
+      "train_samples_per_second": 19.371,
+      "train_steps_per_second": 0.151
+    }
+  ],
+  "max_steps": 15400,
+  "num_train_epochs": 100,
+  "total_flos": 2.652008062738907e+20,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e07ae4abc8c48013db20dc56c875f2a1b7115ee1ed5e58dc64886f0b18aec42a
+size 3055

vocab.json

@@ -0,0 +1,33 @@
+{
+  "[PAD]": 30,
+  "[UNK]": 29,
+  "a": 1,
+  "b": 2,
+  "c": 3,
+  "d": 4,
+  "e": 5,
+  "f": 6,
+  "g": 7,
+  "h": 8,
+  "i": 9,
+  "j": 10,
+  "k": 11,
+  "l": 12,
+  "m": 13,
+  "n": 14,
+  "o": 15,
+  "p": 16,
+  "q": 17,
+  "r": 18,
+  "s": 19,
+  "t": 20,
+  "u": 21,
+  "v": 22,
+  "w": 23,
+  "x": 24,
+  "y": 25,
+  "z": 26,
+  "|": 0,
+  "‘": 27,
+  "’": 28
+}

with_ngram_LM.ipynb

@@ -0,0 +1,350 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "id": "072d16f1",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "/workspace/xls-r-uzbek-cv8\n"
+     ]
+    }
+   ],
+   "source": [
+    "%cd ~/xls-r-uzbek-cv8"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 10,
+   "id": "12382315",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "\u001b[33mWARNING: Ignoring invalid distribution -ransformers (/opt/conda/lib/python3.8/site-packages)\u001b[0m\u001b[33m\n",
+      "\u001b[0m\u001b[33mWARNING: Ignoring invalid distribution -ip (/opt/conda/lib/python3.8/site-packages)\u001b[0m\u001b[33m\n",
+      "\u001b[0m\u001b[33mWARNING: Ignoring invalid distribution - (/opt/conda/lib/python3.8/site-packages)\u001b[0m\u001b[33m\n",
+      "\u001b[0m\u001b[33mWARNING: Ignoring invalid distribution -ransformers (/opt/conda/lib/python3.8/site-packages)\u001b[0m\u001b[33m\n",
+      "\u001b[0m\u001b[33mWARNING: Ignoring invalid distribution -ip (/opt/conda/lib/python3.8/site-packages)\u001b[0m\u001b[33m\n",
+      "\u001b[0m\u001b[33mWARNING: Ignoring invalid distribution - (/opt/conda/lib/python3.8/site-packages)\u001b[0m\u001b[33m\n",
+      "\u001b[0mCollecting https://github.com/kpu/kenlm/archive/master.zip (from -r requirements.txt (line 10))\n",
+      "  Using cached https://github.com/kpu/kenlm/archive/master.zip (541 kB)\n",
+      "  Preparing metadata (setup.py) ... \u001b[?25ldone\n",
+      "\u001b[?25hRequirement already satisfied: unidecode in /opt/conda/lib/python3.8/site-packages (from -r requirements.txt (line 1)) (1.3.2)\n",
+      "Collecting tensorboard\n",
+      "  Using cached tensorboard-2.8.0-py3-none-any.whl (5.8 MB)\n",
+      "Requirement already satisfied: torch in /opt/conda/lib/python3.8/site-packages (from -r requirements.txt (line 3)) (1.10.2)\n",
+      "Requirement already satisfied: torchaudio in /opt/conda/lib/python3.8/site-packages (from -r requirements.txt (line 4)) (0.10.2)\n",
+      "Requirement already satisfied: jiwer~=2.3.0 in /opt/conda/lib/python3.8/site-packages (from -r requirements.txt (line 5)) (2.3.0)\n",
+      "Requirement already satisfied: soundfile~=0.10.3 in /opt/conda/lib/python3.8/site-packages (from -r requirements.txt (line 6)) (0.10.3.post1)\n",
+      "Collecting transformers~=4.16.2\n",
+      "  Using cached transformers-4.16.2-py3-none-any.whl (3.5 MB)\n",
+      "Collecting datasets~=1.18.3\n",
+      "  Using cached datasets-1.18.3-py3-none-any.whl (311 kB)\n",
+      "Requirement already satisfied: pyctcdecode in /opt/conda/lib/python3.8/site-packages (from -r requirements.txt (line 9)) (0.3.0)\n",
+      "Requirement already satisfied: protobuf>=3.6.0 in /opt/conda/lib/python3.8/site-packages (from tensorboard->-r requirements.txt (line 2)) (3.19.4)\n",
+      "Requirement already satisfied: tensorboard-plugin-wit>=1.6.0 in /opt/conda/lib/python3.8/site-packages (from tensorboard->-r requirements.txt (line 2)) (1.8.1)\n",
+      "Collecting google-auth-oauthlib<0.5,>=0.4.1\n",
+      "  Using cached google_auth_oauthlib-0.4.6-py2.py3-none-any.whl (18 kB)\n",
+      "Requirement already satisfied: google-auth<3,>=1.6.3 in /opt/conda/lib/python3.8/site-packages (from tensorboard->-r requirements.txt (line 2)) (2.6.0)\n",
+      "Requirement already satisfied: numpy>=1.12.0 in /opt/conda/lib/python3.8/site-packages (from tensorboard->-r requirements.txt (line 2)) (1.19.2)\n",
+      "Requirement already satisfied: setuptools>=41.0.0 in /opt/conda/lib/python3.8/site-packages (from tensorboard->-r requirements.txt (line 2)) (50.3.1.post20201107)\n",
+      "Requirement already satisfied: requests<3,>=2.21.0 in /opt/conda/lib/python3.8/site-packages (from tensorboard->-r requirements.txt (line 2)) (2.24.0)\n",
+      "Requirement already satisfied: markdown>=2.6.8 in /opt/conda/lib/python3.8/site-packages (from tensorboard->-r requirements.txt (line 2)) (3.3.6)\n",
+      "Requirement already satisfied: grpcio>=1.24.3 in /opt/conda/lib/python3.8/site-packages (from tensorboard->-r requirements.txt (line 2)) (1.43.0)\n",
+      "Requirement already satisfied: wheel>=0.26 in /opt/conda/lib/python3.8/site-packages (from tensorboard->-r requirements.txt (line 2)) (0.35.1)\n",
+      "Requirement already satisfied: absl-py>=0.4 in /opt/conda/lib/python3.8/site-packages (from tensorboard->-r requirements.txt (line 2)) (1.0.0)\n",
+      "Requirement already satisfied: werkzeug>=0.11.15 in /opt/conda/lib/python3.8/site-packages (from tensorboard->-r requirements.txt (line 2)) (2.0.2)\n",
+      "Requirement already satisfied: tensorboard-data-server<0.7.0,>=0.6.0 in /opt/conda/lib/python3.8/site-packages (from tensorboard->-r requirements.txt (line 2)) (0.6.1)\n",
+      "Requirement already satisfied: typing-extensions in /opt/conda/lib/python3.8/site-packages (from torch->-r requirements.txt (line 3)) (4.0.1)\n",
+      "Requirement already satisfied: python-Levenshtein==0.12.2 in /opt/conda/lib/python3.8/site-packages (from jiwer~=2.3.0->-r requirements.txt (line 5)) (0.12.2)\n",
+      "Requirement already satisfied: cffi>=1.0 in /opt/conda/lib/python3.8/site-packages (from soundfile~=0.10.3->-r requirements.txt (line 6)) (1.14.3)\n",
+      "Requirement already satisfied: filelock in /opt/conda/lib/python3.8/site-packages (from transformers~=4.16.2->-r requirements.txt (line 7)) (3.0.12)\n",
+      "Requirement already satisfied: regex!=2019.12.17 in /opt/conda/lib/python3.8/site-packages (from transformers~=4.16.2->-r requirements.txt (line 7)) (2022.1.18)\n",
+      "Requirement already satisfied: sacremoses in /opt/conda/lib/python3.8/site-packages (from transformers~=4.16.2->-r requirements.txt (line 7)) (0.0.47)\n",
+      "Requirement already satisfied: packaging>=20.0 in /opt/conda/lib/python3.8/site-packages (from transformers~=4.16.2->-r requirements.txt (line 7)) (21.3)\n",
+      "Requirement already satisfied: tokenizers!=0.11.3,>=0.10.1 in /opt/conda/lib/python3.8/site-packages (from transformers~=4.16.2->-r requirements.txt (line 7)) (0.11.4)\n",
+      "Requirement already satisfied: huggingface-hub<1.0,>=0.1.0 in /opt/conda/lib/python3.8/site-packages (from transformers~=4.16.2->-r requirements.txt (line 7)) (0.4.0)\n",
+      "Requirement already satisfied: pyyaml>=5.1 in /opt/conda/lib/python3.8/site-packages (from transformers~=4.16.2->-r requirements.txt (line 7)) (5.4.1)\n",
+      "Requirement already satisfied: tqdm>=4.27 in /opt/conda/lib/python3.8/site-packages (from transformers~=4.16.2->-r requirements.txt (line 7)) (4.62.3)\n",
+      "Requirement already satisfied: multiprocess in /opt/conda/lib/python3.8/site-packages (from datasets~=1.18.3->-r requirements.txt (line 8)) (0.70.12.2)\n",
+      "Requirement already satisfied: dill in /opt/conda/lib/python3.8/site-packages (from datasets~=1.18.3->-r requirements.txt (line 8)) (0.3.4)\n",
+      "Requirement already satisfied: pandas in /opt/conda/lib/python3.8/site-packages (from datasets~=1.18.3->-r requirements.txt (line 8)) (1.4.0)\n",
+      "Requirement already satisfied: aiohttp in /opt/conda/lib/python3.8/site-packages (from datasets~=1.18.3->-r requirements.txt (line 8)) (3.8.1)\n",
+      "Requirement already satisfied: xxhash in /opt/conda/lib/python3.8/site-packages (from datasets~=1.18.3->-r requirements.txt (line 8)) (2.0.2)\n",
+      "Requirement already satisfied: pyarrow!=4.0.0,>=3.0.0 in /opt/conda/lib/python3.8/site-packages (from datasets~=1.18.3->-r requirements.txt (line 8)) (6.0.1)\n",
+      "Requirement already satisfied: fsspec[http]>=2021.05.0 in /opt/conda/lib/python3.8/site-packages (from datasets~=1.18.3->-r requirements.txt (line 8)) (2022.1.0)\n",
+      "Requirement already satisfied: pygtrie<3.0,>=2.1 in /opt/conda/lib/python3.8/site-packages (from pyctcdecode->-r requirements.txt (line 9)) (2.4.2)\n",
+      "Requirement already satisfied: hypothesis<7,>=6.14 in /opt/conda/lib/python3.8/site-packages (from pyctcdecode->-r requirements.txt (line 9)) (6.36.1)\n",
+      "Requirement already satisfied: six in /opt/conda/lib/python3.8/site-packages (from absl-py>=0.4->tensorboard->-r requirements.txt (line 2)) (1.15.0)\n",
+      "Requirement already satisfied: pycparser in /opt/conda/lib/python3.8/site-packages (from cffi>=1.0->soundfile~=0.10.3->-r requirements.txt (line 6)) (2.20)\n",
+      "Requirement already satisfied: rsa<5,>=3.1.4 in /opt/conda/lib/python3.8/site-packages (from google-auth<3,>=1.6.3->tensorboard->-r requirements.txt (line 2)) (4.8)\n",
+      "Requirement already satisfied: cachetools<6.0,>=2.0.0 in /opt/conda/lib/python3.8/site-packages (from google-auth<3,>=1.6.3->tensorboard->-r requirements.txt (line 2)) (5.0.0)\n",
+      "Requirement already satisfied: pyasn1-modules>=0.2.1 in /opt/conda/lib/python3.8/site-packages (from google-auth<3,>=1.6.3->tensorboard->-r requirements.txt (line 2)) (0.2.8)\n",
+      "Requirement already satisfied: requests-oauthlib>=0.7.0 in /opt/conda/lib/python3.8/site-packages (from google-auth-oauthlib<0.5,>=0.4.1->tensorboard->-r requirements.txt (line 2)) (1.3.1)\n",
+      "Requirement already satisfied: sortedcontainers<3.0.0,>=2.1.0 in /opt/conda/lib/python3.8/site-packages (from hypothesis<7,>=6.14->pyctcdecode->-r requirements.txt (line 9)) (2.4.0)\n",
+      "Requirement already satisfied: attrs>=19.2.0 in /opt/conda/lib/python3.8/site-packages (from hypothesis<7,>=6.14->pyctcdecode->-r requirements.txt (line 9)) (21.4.0)\n",
+      "Requirement already satisfied: importlib-metadata>=4.4 in /opt/conda/lib/python3.8/site-packages (from markdown>=2.6.8->tensorboard->-r requirements.txt (line 2)) (4.10.1)\n",
+      "Requirement already satisfied: pyparsing!=3.0.5,>=2.0.2 in /opt/conda/lib/python3.8/site-packages (from packaging>=20.0->transformers~=4.16.2->-r requirements.txt (line 7)) (3.0.7)\n",
+      "Requirement already satisfied: chardet<4,>=3.0.2 in /opt/conda/lib/python3.8/site-packages (from requests<3,>=2.21.0->tensorboard->-r requirements.txt (line 2)) (3.0.4)\n",
+      "Requirement already satisfied: idna<3,>=2.5 in /opt/conda/lib/python3.8/site-packages (from requests<3,>=2.21.0->tensorboard->-r requirements.txt (line 2)) (2.10)\n",
+      "Requirement already satisfied: urllib3!=1.25.0,!=1.25.1,<1.26,>=1.21.1 in /opt/conda/lib/python3.8/site-packages (from requests<3,>=2.21.0->tensorboard->-r requirements.txt (line 2)) (1.25.11)\n",
+      "Requirement already satisfied: certifi>=2017.4.17 in /opt/conda/lib/python3.8/site-packages (from requests<3,>=2.21.0->tensorboard->-r requirements.txt (line 2)) (2020.12.5)\n",
+      "Requirement already satisfied: yarl<2.0,>=1.0 in /opt/conda/lib/python3.8/site-packages (from aiohttp->datasets~=1.18.3->-r requirements.txt (line 8)) (1.7.2)\n",
+      "Requirement already satisfied: async-timeout<5.0,>=4.0.0a3 in /opt/conda/lib/python3.8/site-packages (from aiohttp->datasets~=1.18.3->-r requirements.txt (line 8)) (4.0.2)\n",
+      "Requirement already satisfied: aiosignal>=1.1.2 in /opt/conda/lib/python3.8/site-packages (from aiohttp->datasets~=1.18.3->-r requirements.txt (line 8)) (1.2.0)\n",
+      "Requirement already satisfied: multidict<7.0,>=4.5 in /opt/conda/lib/python3.8/site-packages (from aiohttp->datasets~=1.18.3->-r requirements.txt (line 8)) (6.0.2)\n",
+      "Requirement already satisfied: frozenlist>=1.1.1 in /opt/conda/lib/python3.8/site-packages (from aiohttp->datasets~=1.18.3->-r requirements.txt (line 8)) (1.3.0)\n",
+      "Requirement already satisfied: charset-normalizer<3.0,>=2.0 in /opt/conda/lib/python3.8/site-packages (from aiohttp->datasets~=1.18.3->-r requirements.txt (line 8)) (2.0.10)\n",
+      "Requirement already satisfied: python-dateutil>=2.8.1 in /opt/conda/lib/python3.8/site-packages (from pandas->datasets~=1.18.3->-r requirements.txt (line 8)) (2.8.2)\n",
+      "Requirement already satisfied: pytz>=2020.1 in /opt/conda/lib/python3.8/site-packages (from pandas->datasets~=1.18.3->-r requirements.txt (line 8)) (2021.1)\n",
+      "Requirement already satisfied: joblib in /opt/conda/lib/python3.8/site-packages (from sacremoses->transformers~=4.16.2->-r requirements.txt (line 7)) (1.1.0)\n",
+      "Requirement already satisfied: click in /opt/conda/lib/python3.8/site-packages (from sacremoses->transformers~=4.16.2->-r requirements.txt (line 7)) (8.0.3)\n",
+      "Requirement already satisfied: zipp>=0.5 in /opt/conda/lib/python3.8/site-packages (from importlib-metadata>=4.4->markdown>=2.6.8->tensorboard->-r requirements.txt (line 2)) (3.7.0)\n",
+      "Requirement already satisfied: pyasn1<0.5.0,>=0.4.6 in /opt/conda/lib/python3.8/site-packages (from pyasn1-modules>=0.2.1->google-auth<3,>=1.6.3->tensorboard->-r requirements.txt (line 2)) (0.4.8)\n",
+      "Requirement already satisfied: oauthlib>=3.0.0 in /opt/conda/lib/python3.8/site-packages (from requests-oauthlib>=0.7.0->google-auth-oauthlib<0.5,>=0.4.1->tensorboard->-r requirements.txt (line 2)) (3.2.0)\n",
+      "Building wheels for collected packages: kenlm\n",
+      "  Building wheel for kenlm (setup.py) ... \u001b[?25ldone\n",
+      "\u001b[?25h  Created wheel for kenlm: filename=kenlm-0.0.0-cp38-cp38-linux_x86_64.whl size=2348591 sha256=d5c8e5430d89f59ddde39bc78aec471c1e66ef43b6cde792711b2e97d7b8b9dc\n",
+      "  Stored in directory: /tmp/pip-ephem-wheel-cache-hhcfnszu/wheels/ff/08/4e/a3ddc0e786e0f3c1fcd2e7a82c4324c02fc3ae2638471406d2\n",
+      "Successfully built kenlm\n",
+      "\u001b[33mWARNING: Ignoring invalid distribution -ransformers (/opt/conda/lib/python3.8/site-packages)\u001b[0m\u001b[33m\n",
+      "\u001b[0m\u001b[33mWARNING: Ignoring invalid distribution -ip (/opt/conda/lib/python3.8/site-packages)\u001b[0m\u001b[33m\n",
+      "\u001b[0m\u001b[33mWARNING: Ignoring invalid distribution - (/opt/conda/lib/python3.8/site-packages)\u001b[0m\u001b[33m\n",
+      "\u001b[0mInstalling collected packages: kenlm, transformers, google-auth-oauthlib, tensorboard, datasets\n",
+      "\u001b[33m  WARNING: The script transformers-cli is installed in '/workspace/.local/bin' which is not on PATH.\n",
+      "  Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.\u001b[0m\u001b[33m\n",
+      "\u001b[0m\u001b[33m  WARNING: The script google-oauthlib-tool is installed in '/workspace/.local/bin' which is not on PATH.\n",
+      "  Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.\u001b[0m\u001b[33m\n",
+      "\u001b[0m\u001b[33m  WARNING: The script tensorboard is installed in '/workspace/.local/bin' which is not on PATH.\n",
+      "  Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.\u001b[0m\u001b[33m\n",
+      "\u001b[0m\u001b[33m  WARNING: The script datasets-cli is installed in '/workspace/.local/bin' which is not on PATH.\n",
+      "  Consider adding this directory to PATH or, if you prefer to suppress this warning, use --no-warn-script-location.\u001b[0m\u001b[33m\n",
+      "\u001b[0mSuccessfully installed datasets-1.18.3 google-auth-oauthlib-0.4.6 kenlm-0.0.0 tensorboard-2.8.0 transformers-4.16.2\n",
+      "\u001b[33mWARNING: Ignoring invalid distribution -ransformers (/opt/conda/lib/python3.8/site-packages)\u001b[0m\u001b[33m\n",
+      "\u001b[0m\u001b[33mWARNING: Ignoring invalid distribution -ip (/opt/conda/lib/python3.8/site-packages)\u001b[0m\u001b[33m\n",
+      "\u001b[0m\u001b[33mWARNING: Ignoring invalid distribution - (/opt/conda/lib/python3.8/site-packages)\u001b[0m\u001b[33m\n",
+      "\u001b[0m"
+     ]
+    }
+   ],
+   "source": [
+    "!python -m pip install -r requirements.txt --user"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 14,
+   "id": "3969d63a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from transformers import AutoFeatureExtractor, AutoTokenizer, pipeline\n",
+    "from datasets import Audio, Dataset, DatasetDict, load_dataset, load_metric\n",
+    "\n",
+    "import re\n",
+    "import string\n",
+    "import unidecode"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 12,
+   "id": "daff17fd",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Reusing dataset common_voice (/workspace/.cache/huggingface/datasets/mozilla-foundation___common_voice/uz/8.0.0/b8bc4d453193c06a43269b46cd87f075c70f152ac963b7f28f7a2760c45ec3e8)\n"
+     ]
+    },
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "a8aad37a859241ff81ac932edc204bf8",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "  0%|          | 0/5 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "text/plain": [
+       "DatasetDict({\n",
+       "    train: Dataset({\n",
+       "        features: ['client_id', 'path', 'audio', 'sentence', 'up_votes', 'down_votes', 'age', 'gender', 'accent', 'locale', 'segment'],\n",
+       "        num_rows: 39456\n",
+       "    })\n",
+       "    test: Dataset({\n",
+       "        features: ['client_id', 'path', 'audio', 'sentence', 'up_votes', 'down_votes', 'age', 'gender', 'accent', 'locale', 'segment'],\n",
+       "        num_rows: 11598\n",
+       "    })\n",
+       "    validation: Dataset({\n",
+       "        features: ['client_id', 'path', 'audio', 'sentence', 'up_votes', 'down_votes', 'age', 'gender', 'accent', 'locale', 'segment'],\n",
+       "        num_rows: 10849\n",
+       "    })\n",
+       "    other: Dataset({\n",
+       "        features: ['client_id', 'path', 'audio', 'sentence', 'up_votes', 'down_votes', 'age', 'gender', 'accent', 'locale', 'segment'],\n",
+       "        num_rows: 119461\n",
+       "    })\n",
+       "    invalidated: Dataset({\n",
+       "        features: ['client_id', 'path', 'audio', 'sentence', 'up_votes', 'down_votes', 'age', 'gender', 'accent', 'locale', 'segment'],\n",
+       "        num_rows: 11276\n",
+       "    })\n",
+       "})"
+      ]
+     },
+     "execution_count": 12,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "dataset_dict = load_dataset(\"mozilla-foundation/common_voice_8_0\", \"uz\", use_auth_token=True)\n",
+    "dataset_dict"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 17,
+   "id": "c22aee32",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "chars_to_ignore_regex=f\"[{re.escape(string.punctuation)}]\" \n",
+    "\n",
+    "def remove_special_characters(batch):\n",
+    "    batch[\"text\"] = re.sub(\n",
+    "        chars_to_ignore_regex, \n",
+    "        \"\", \n",
+    "        re.sub(\"['`´]\", \"’\",   # elsewhere probably meant as glottal stop\n",
+    "               re.sub(\"([og])['`´]\", \"\\g<1>‘\",  # after o/g indicate modified char\n",
+    "                      unidecode.unidecode(batch[\"sentence\"]).lower()\n",
+    "                     )\n",
+    "              )\n",
+    "    ) + \" \"\n",
+    "    return batch"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 18,
+   "id": "f28dc522",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "4b8d2f0df8ea46bdaee2c94996583c5e",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "0ex [00:00, ?ex/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "dataset = dataset_dict[\"train\"].map(remove_special_characters, remove_columns=dataset_dict[\"train\"].column_names)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 23,
+   "id": "38e02d29",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "      0  244494 2030240 uz_cv8_train.txt\n"
+     ]
+    }
+   ],
+   "source": [
+    "text_data = \"uz_cv8_train.txt\"\n",
+    "with open(text_data, \"w\") as fs:\n",
+    "    fs.write(\" \".join(dataset[\"text\"]))\n",
+    "\n",
+    "!wc $text_data"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 26,
+   "id": "7b3d70f0",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "--2022-02-07 03:18:36--  https://kheafield.com/code/kenlm.tar.gz\n",
+      "Resolving kheafield.com (kheafield.com)... 35.196.63.85\n",
+      "Connecting to kheafield.com (kheafield.com)|35.196.63.85|:443... connected.\n",
+      "HTTP request sent, awaiting response... 200 OK\n",
+      "Length: 491090 (480K) [application/x-gzip]\n",
+      "Saving to: ‘STDOUT’\n",
+      "\n",
+      "-                   100%[===================>] 479.58K  2.31MB/s    in 0.2s    \n",
+      "\n",
+      "2022-02-07 03:18:37 (2.31 MB/s) - written to stdout [491090/491090]\n",
+      "\n",
+      "/bin/bash: line 1: cmake: command not found\n"
+     ]
+    }
+   ],
+   "source": [
+    "!wget -O - https://kheafield.com/code/kenlm.tar.gz | tar xz\n",
+    "!mkdir kenlm/build && cd kenlm/build && cmake .. && make -j2\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "65118a69",
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "Python 3 (ipykernel)",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.8.8"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}